Comparative analysis of deep decoder and U-Net architectures for propagation-based phase reconstruction from intensity measurements using untrained neural networks

Untrained neural networks (UNNs) have recently emerged as a promising approach for solving inverse problems without the need for large-scale datasets or extensive training. In this work, we systematically investigate the performance of UNNs in the context of phase retrieval from free-space propagation of intensity. We focus on two distinct architectures: the deep decoder and U-Net, evaluating their ability to reconstruct phase information from intensity measurements. Through simulation and experiment, we compare the reconstruction robustness of both architectures under various noise levels and propagation distances. Our findings highlight the strengths and limitations of each architecture, offering insights into the design of effective UNN-based phase retrieval algorithms and guiding future work toward more efficient imaging solutions.